Due to the rapid development of science and technology, many schools and teachers use the e-learning platform as a teaching aid tool. Therefore, the platform records a large number of students' learning behaviors. How to make full use of these data to improve the effectiveness of the curriculum is what most scholars are concerned about. We hope to use a large amount of recorded data on the platform to predict students' learning behaviors through machine learning methods to meet the individual needs of different students. We can also take remedial measures for students at risk, with the ability to assess the learning status, predicting the future learning status and performance from the student's learning process, and then providing teachers with additional information to help improve the student's learning status. So we can turn the e-learning platform from playing a passive role on data-access-record to an active one with evaluation-caution-transformation. There are three major challenges in predicting student performance. The first is that student records are often high-dimensional data, which makes the prediction effect poor. The second problem is that the data is time-series data, so the order of the data must be considered. In order, the third is that in most cases, we only have partial information, so it is challenging to use partial data to make accurate predictions. In this paper, we will use phase space partition to split high-dimensional data, and use the symbolic label to represent partitions. These symbolic label sequences can be regarded as discrete sequences. Finally, HiSeqGAN, the neural network of sequence synthesis is used to generate a large amount of data, and use Label Error to calculate the symbol label distance between the generated data and the real data to predict the future performance of students. Then use the methods mentioned above to solve the three major challenges.